// directly, instead of requiring us to serialize to a GraphDef and
// call Session::Extend().
bool ExtendSessionGraphHelper(TF_Session* session, TF_Status* status) {
  if (session->graph != nullptr) {
    // Take the graph lock before the session lock to avoid deadlock. This is
    // safe since session->graph does not change.
    session->graph->mu.lock();
    mutex_lock session_lock(session->mu);
    const Graph& graph = session->graph->graph;

                             serialized_update.size(), status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // The device of var_handle0 is local device which is the same before and
  // after cluster update. Remove resource with valid device should succeed.
  TFE_Op* op = TFE_NewOp(ctx, "DestroyResourceOp", status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

      immediate_execution_ctx_.get(), {x.get()}, UseFunction()));
}

TEST_P(CppGradients, TestMatMulGrad) {
  // TODO(vnvo2409): Figure out why `gradient_checker` does not work very
  // well with `MatMul` and remove `TestMatMul*` in
  // `mnist_gradients_test` when done.
  GTEST_SKIP();

  float A_vals[] = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f};
  int64_t A_dims[] = {3, 3};

  EXPECT_EQ(TF_GetCode(status.get()), TF_INVALID_ARGUMENT);
  EXPECT_THAT(TF_Message(status.get()), HasSubstr("assertion failed"));

  // Note that future collectives with the same context do not work at the
  // moment; once canceled, the collective executor requires the program to be
  // restarted / context to be reset.
}

TEST(PARALLEL_DEVICE_LIB, TestDifferentShapes) {

// GraphDef type. These functions must be deleted by calling TF_DeleteFunction.
//
// If `mutate_proto_func` is non-NULL, run it over each FunctionDef proto,
// before creating a TF_Function out of the possibly mutated proto.
static std::vector<UniqueFuncPtr> CreateFunctionsFromTextProto(
    const char* text_proto,
    std::function<void(FunctionDef*)>* mutate_proto_func, TF_Status* status) {

  tensorflow::mutex execution_mutex_;
  ExecutionState execution_state_ TF_GUARDED_BY(execution_mutex_) =
      ExecutionState::kIdle;
  // Tells the worker thread that there is new work.
  tensorflow::condition_variable start_execute_;
  // The worker thread notifies that work has finished.
  tensorflow::condition_variable finished_execute_;
  // Notifies a StartExecute that the previous Join has finished.
  tensorflow::condition_variable finished_join_;

  int num_dims = TF_NumDims(t);
  for (int i = 0; i < num_dims; i++) {
    out_dims[i] = TF_Dim(t, i);
  }
}

// Runs model as is if output is a scalar,
// else sums the output tensor before returning.
Status RunAndMaybeSum(AbstractContext* ctx, Model forward,
                      absl::Span<AbstractTensorHandle* const> inputs,
                      absl::Span<AbstractTensorHandle*> outputs,

  TFE_TensorHandle* h1 = TestVariable(ctx, 2.0, task2_name);
  TFE_TensorHandle* h2 = TestVariable(ctx, 3.0, task0_name);

  // Add a sync point to make sure that variables have been initialized
  // before the function execution starts.
  TFE_ContextAsyncWait(ctx, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // Pack 3 variable handles into one TFE_TensorHandle.

#define ASSERT_TF_OK(x) ASSERT_EQ(TF_OK, TF_GetCode(x)) << TF_Message(x)
#define EXPECT_TF_OK(x) EXPECT_EQ(TF_OK, TF_GetCode(x)) << TF_Message(x)

static const char* content = "abcdefghijklmnopqrstuvwxyz1234567890";
// We will work with content_view instead of content.
static const absl::string_view content_view = content;

namespace gcs = google::cloud::storage;

static std::string InitializeTmpDir() {

      TF_SetStatus(status, TF_OUT_OF_RANGE, std::move(os).str().c_str());
      return total_bytes_transferred;
    }
    auto begin = data.begin();
    if (offset > pos) {
      // The block begins before the slice we're reading.
      begin += offset - pos;
    }
    auto end = data.end();
    if (pos + data.size() > offset + n) {
      // The block extends past the end of the slice we're reading.